Driving Scene Segmentation

Performing Segmentation on Indian Roads.

Self-driving cars also known as Autonomous driving cars are key innovation which has revolutionized the automobile industry. Autonomous driving is a complex task which requires precise understanding of the environment. In order to get precise pixel wise information of the driving scenes, semantic segmentation is used.

Approach

DataSet

The environment of Indian roads are unconstrained and unconstrained. So, inorder to perform segmentation on the Indian roads. I have used Indian Driving Dataset(IDD). This dataset consist of 16,000 images with 21 classes collected from 182 drive sequences.

Model

To Perform Segmentation, I have used slightly modified U-Net. I have added BatchNorm after each conv layer and instead of upconv, I have used convtranspose layer.

Training

I have used JaccardLoss from segmentation_models_pytorch library. Optimizer Used Adam with a Learning rate of 1e-3. Trained for 100 epochs. Dice Score achieved after Training 80.04

To run the project

git clone https://github.com/mohan-gupta/driving-scene-segmentation.git  # clone
cd driving-scene-segmentation
pip install -r requirements.txt  # install
cd app
streamlit run streamlit.py  #run